Measure the cost of task grant status checks

[jeremi]

Follow-up to #1029. Both behaviours fail closed, which is intended; this is about load and latency, not correctness.

• BREG check_task_authority calls the Casework status endpoint while holding the proposal row lock and a database connection, with a 5 s timeout. A slow Casework can hold both for up to 5 s per grant-bound write.

• Casework task_status does a live source read on every call, so each agent write costs a source round trip, and a source outage stops agent writes.

• Load test grant-bound BREG writes with Casework and the source slowed down, and record pool use and latency

• Based on the result, decide whether the status call moves outside the lock, or Casework reuses source context within the grant lifetime, or nothing changes